Automating the bottleneck: a scalable deep learning pipeline for phytolith detection, imaging, and curation across...

Deep learning offers powerful tools for automating phytolith classification. Although foundational models have enabled image classification with little or no task specific training, specialized domains such as phytolith images still require a labor-intensive process to acquire labelled training data. This article presents a novel, scalable imaging workflow that addresses this bottleneck by integrating low-magnification slide scanning with artificial intelligence assisted phytolith detection and targeted high-magnification 3D Z-stack capture optimized for both multi-cell and single-cell phytoliths. Automated pre-processing enables imaging of 16 slides per week generating 100's of curated phytolith training images per slide. Developed in collaboration with Nikon Imaging Specialists and validated on 34 comparative grass and rush taxa, the system combines expert-guided sampling with customized automated acquisition routines to capture, segment, tile and curate images for deep learning input. Key workflow optimizations, including hardware upgrades, asynchronous file conversion, and iterative refinement of the artificial intelligence detection model, achieved a 350% improvement in extended depth of field rendering speed and a 55,450% improvement in TIFF export speed, enabling sustained high-throughput acquisition without performance degradation. The workflow was designed to optimize imaging of mixed multi-cell and single-cell phytolith assemblages and is currently generating training data for DL models targeting wild versus domesticated wheat discrimination in addition to grass species identification from multi-cell husk phytolith morphology. Beyond phytoliths, the pipeline is modular, replicable across microscopy platforms, and readily adaptable to other plant microfossils and microstructures, including pollen, starch grains and other silica bodies, through targeted artificial intelligence retraining, making large-scale machine learning applications in plant science, paleoethnobotany and related fields more accessible and reproducible.